CERN and Mistral plan one AI model to read every LHC collision

CERN and Mistral have signed a statement of intent for a long-term partnership, according to CERN’s announcement. Its centerpiece is a single Large Physics Model, billed as the first foundation model for collider physics, trained on billions of simulated Large Hadron Collider collisions.

One model for the whole reconstruction chain

CERN dated the announcement 15 October 2026. It describes the deal as a statement of intent to set up a strategic partnership on foundation models for scientific discovery. CERN brings expertise in fundamental physics, scientific data and large-scale computing. Mistral, a European developer of large AI models, brings experience in training them at scale. The first focus is high-energy physics, though CERN says the approach could eventually serve other fields.

The model is to be trained on simulated proton-proton collisions from the LHC. CERN says it should handle several jobs at once: reconstructing collisions, identifying particles, flagging unexpected signals and evaluating future detector designs.

The problem it targets is how LHC data is processed today. CERN explains that reconstruction relies on a chain of specialized algorithms, which step by step turn detector signals into the particles and other objects used in physics analysis. Machine learning already improves many of those steps. The idea is that a foundation model could learn across the entire chain instead of being built task by task.

CERN’s Chief Information Officer, Enrica Porcari, said that high-energy physics offers an environment where AI can be developed and tested “with exceptional rigour” and that the aim is capabilities “that can serve the whole physics community and be shared”. Mistral’s Johannes Brandstetter, Vice-President of AI for Engineering, said the companies have the chance to push the frontiers of high-energy physics.

Why the data volume makes this timely

CERN says it holds about one exabyte, or one million terabytes, of LHC experimental data. It also says the High-Luminosity LHC, expected to start around 2030, will raise the number of collisions tenfold.

Outside reporting adds context. Trade outlet Converge Digest, covering CERN’s December 2025 storage milestone, reported that the exabyte reflects more than 15 years of LHC operation and that much of it sits on magnetic tape. It also reported that the upgraded collider is expected to generate roughly ten times more data than today’s machine, which matches CERN’s own figure. Mistral, for its part, describes itself as a Paris-headquartered lab building open-weight models.

That scale is the case for a shared model. If analysis chains have to be retuned for each task and each detector, a tenfold rise in collisions multiplies the work. A pretrained model that teams could adapt would, in principle, spread the training cost across many analyses.

What would confirm the claims

The announcement is a statement of intent, not a product. It gives no budget, no timeline, no model size, no compute commitment and no word on whether the weights or training data will be released. Porcari’s remark that the capabilities should be shared points toward openness, but nothing in the text commits to it. Readers should treat the Large Physics Model as a plan.

Several things will show whether it is real:

• A formal agreement. A signed partnership with named resources, rather than a statement of intent, would show how much each side is committing.
• Published benchmarks. The model is trained on simulation, so the open question is how well it performs on real detector data, and against the specialized algorithms the experiments use today. CERN makes no performance claims yet.
• Access terms. Whether outside physicists can use or fine-tune the model will decide whether it serves the wider community CERN mentions.
• Timing against the HL-LHC. A usable model before the upgraded collider starts around 2030 would matter most, but no date has been given.

For now, the announcement establishes the aim and the partners. All the claims about what the model could do come from CERN and Mistral themselves and have not been independently tested. Foundation models have worked well in language and images, and whether the same recipe holds for collider events, with their strict demands on precision, is what this project will have to show.